Frank Robotics Lab中级课程，教你通过MCP协议将本地LLM接入ROS 2机器人，在Ubuntu 24.04仿真环境中用Ollama量化测试任务成功率，并实现地理围栏、限速与防注入等安全约束。
原始标题：LLM Agents for ROS 2 Robots: Local Models, MCP, Safety
这门由 Frank Robotics Lab 推出的中级课程专注于通过 MCP 协议将本地大语言模型（LLM）连接到 ROS 2 机器人，并进行量化测试与安全约束。课程完全在 Ubuntu 24.04 (ROS 2 Jazzy) 仿真环境中运行，使用 Ollama 部署本地轻量化模型（如 qwen2.5:3b），拒绝空洞的 Demo 展示，全程用真实机器人位姿数据和统计学区间来量化评估智能体的任务成功率。
课程核心内容涵盖 MCP 机器人工具设计、Nav2 导航仿真、服务器端硬安全防御（地理围栏、限速、杀毒开关）以及防提示词注入攻击。学员将通过 36 节带有“经典错误对比与自动化校验脚本”的课程，掌握如何构建安全、可测量且具备防御能力的嵌入式机器人 LLM 智能体控制闭环。
Published 10/2026
Created by Frank Robotics Lab
MP4 | Video: h264, 1920×1080 | Audio: AAC, 44.1 KHz, 2 Ch
Level: Intermediate | Genre: eLearning | Language: English + subtitle | Duration: 36 Lectures ( 1h 58m ) | Size: 1.8 GB
Connect a local LLM to a ROS 2 robot through MCP, then measure, constrain and test it.
What you’ll learn
⚡ Expose ROS 2 topics, services and actions to a local LLM as MCP tools (mcp 2.x).
⚡ Design tools the model gets right: skills over raw velocity, units, actionable errors.
⚡ Score every run from the robot’s true pose and path, with Wilson intervals, never from the model’s reply.
⚡ Drive Nav2 in a headless simulation through MCP, with cancel and timeouts.
⚡ Enforce safety in the server: geofence, two-sided speed limits, human approval, audit log, kill switch.
⚡ Measure prompt injection through robot data, and defend with typed fields, provenance and action guards.
Requirements
❗ Python, and basic ROS 2: nodes, topics, services, actions.
❗ Ubuntu 24.04 with ROS 2 Jazzy, and Ollama. Every check script runs without a GPU or the model; an NVIDIA GPU makes re-recording model trials faster.
❗ About 6 GB of free disk (measured in lecture 1.1: models, course files, ROS 2, Gazebo and Nav2).
Description
This course contains the use of artificial intelligence. The narration is a synthetic voice; every number on screen comes from a lab you can run.
36 lectures, about 2.0 hours. Each lecture has a folder with the working code, a broken copy with one classic mistake, and a check script that passes only when the fix works and the mistake is caught.
What was measured, with a small local model (qwen2.5:3b) on one laptop GPU. Your numbers may differ by a few trials
✅ Tool design: raw velocity tools reached the target 0 of 20 times; a go_to skill tool, 15 of 20 (95% interval 0.53 to 0.89). Lecture 3.2.
✅ Safety: a keep-out rule written only in the prompt was broken in 13 of 20 trials; the server-side fence, 0 of 20, at a measured cost: it also stopped 7 of 7 targets whose straight path crossed the zone. Lecture 6.1.
✅ Capstone: a 3-waypoint patrol completed 19 of 20 times, scored from the robot’s true path. Lecture 8.1.
✅ Choosing a model: a smaller model completed 0 of 20 missions while saying DONE every time. Lecture 8.2.
What this course does not cover: real robots (everything runs in turtlesim and a simulated TurtleBot), frontier or cloud models, vision language action models, fine-tuning, manipulation with MoveIt, multi-robot fleets.
Who this course is for
⭐ Robotics engineers who want LLM agents on their robots, measured instead of demoed.
⭐ AI engineers who want to connect agents to real control loops safely.
⭐ Not for complete beginners to Python or ROS 2, and not a course about training models.
此处内容需要权限查看
会员免费查看
